Worked example: scoping a warehouse picking dataset
範例
A team wants to sell a grasping dataset for warehouse picking. They can generate simulated grasps cheaply, and they can also run real teleoperated grasps, which cost far more per episode. Assume the task is picking rigid, uniformly shaped boxes from a shelf. Decide what mix of synthetic and real data the offering should contain, and justify it on cost, scale, fidelity, and transfer risk.
示範講解
Step 1 — Identify what the task actually demands. Picking rigid boxes mainly requires recognizing the box, locating it in 3D, and closing the gripper to a secure hold. Contact physics matter, but the objects are simple and consistent.
Step 2 — Ask what synthetic generation is good at here. Because the boxes are uniform and the scenario is narrow, the modeled world can reproduce the relevant geometry and appearance closely. Millions of simulated grasps can be generated across varied positions and lighting at low marginal cost, which real capture could never match on volume.
Step 3 — Ask where the gap bites. The risk is not object recognition; it is contact. Simulated gripper contact and friction are approximations, so a policy tuned only on synthetic grasps may fail on the real gripper's slip behavior. That is exactly the sim-to-real gap for this task.
Step 4 — Use the real captures as the anchor. A smaller set of real teleoperated grasps records the actual gripper's contact behavior and the real sensor characteristics. Those episodes let the team measure the gap and correct for it, rather than guessing.
Step 5 — Combine. The offering pairs large-scale synthetic grasps for coverage with a smaller real set that anchors the contact behavior. The synthetic portion supplies scale and rare placements; the real portion supplies fidelity where the model is weakest.